Two recent developments — one in schema vocabulary design, the other in AI-generated recommendations — point toward the same underlying transition.
On 7 May 2026 Google deprecated FAQ rich results entirely. Three days later, Joost de Valk filed a proposal at schema.org introducing FAQSection as a subtype of WebPageElement, with a structural diagnosis of why FAQPage had been misused at scale: the schema vocabulary itself never had an honest type for a page that contains an FAQ section as part of its content rather than as its primary purpose. Publishers who wanted to mark up that section had two choices — declare FAQPage and misrepresent what the page is, or skip the markup entirely. Most chose misrepresentation, because the rich result reward was sitting on that side of the choice.
Two days after that, on 12 May 2026, Lily Ray published a separate observation on LinkedIn. Across three independent verticals — procurement software, membership management software, and scheduling software — Google AI Overviews appeared to be doing something new with self-promotional “best of” listicles. Ramp’s own “Best Procurement Software” page was cited as a source in the AI Overview. Ramp itself was absent from the recommendation list. Outseta’s “15 Best Membership Management Software” page was cited. Outseta was absent from the recommendations. Rippling’s “8 Best Scheduling Software” page was cited. Rippling was absent. Three verticals, same pattern: the listicle informed the answer, but the publisher of the listicle did not appear in the answer.
These two observations look unrelated. One is about schema vocabulary design at the standards-body level. The other is about how a specific AI surface treats self-referential commercial content. The connection is not obvious.
They are manifestations of the same shift.
What FAQPage misuse and self-promotional listicle citation have in common is that both succeeded for years on a single mechanic: the page declared something about itself, and consuming systems took the declaration at face value. FAQPage declared “this page is primarily an FAQ” and Google rendered a rich result. Ramp declared “these are the best procurement tools, ourselves first” and Google AI Overviews surfaced Ramp on the recommendation list. The mechanism was the same in both cases: self-declaration translated directly into system trust.
That mechanism is breaking down. In the schema case, the breakdown is structural — the vocabulary itself is being challenged by a proposal that would require publishers to describe what their pages actually contain rather than what they wish their pages were classified as. In the AI recommendation case, the breakdown is behavioural — the system continues to use the self-referential content for entity discovery and market intelligence, but it appears to apply a trust discount when synthesising the actual recommendation.
Both developments are early signals of a transition that is going to define how AI systems read the web for the rest of this decade: the shift from syntactic validation to semantic credibility. Schema validators check whether your JSON-LD parses correctly. AI retrieval systems are increasingly evaluating whether what it claims is corroborated. The two are not the same thing. The first is a technical check the publisher controls. The second is a trust judgement the system makes by cross-referencing what your page says against everything else the system knows.
This guide makes the case that semantic credibility is becoming the standard structured data must meet to retain its commercial value. It traces the transition through the FAQSection case study (Part Three), the governance machinery that produces schema standards in the first place (Part Four), and the AI-extraction shift that explains why the validity/credibility distinction matters now in a way it did not five years ago (Parts Five and Six). It introduces the Schema Half-Life Pattern (Part Two) as a predictive framework for which schema types survive the transition and which decay. And it closes (Part Seven) with the strategic implications for any business that depends on AI systems being able to interpret its web presence accurately.
The piece credits Joost de Valk for the schema-vocabulary diagnosis and Lily Ray for the AI Overview observation. The synthesis — that these are manifestations of a single shift, and what that shift implies for the AI-readable web of the next five years — is the contribution this guide is making to the conversation.
Part Two — The Schema Half-Life Pattern
Before getting to the structural analysis, a pattern worth naming.
Schema types have measurable lifespans. Some survive a decade or more and become foundational reference vocabulary. Others arrive with fanfare, get adopted widely, get gamed at scale, and have their rich result rewards withdrawn within a few years. The pattern is consistent enough across the last decade of structured data history to be worth treating as a named framework.
The Schema Half-Life Pattern. Schema types whose value depends on platform rewards tend to decay. Schema types describing stable real-world entities tend to persist.
The Schema Half-Life Pattern is registered as a named framework in the SEO Strategy Ltd Frameworks Register alongside CITATE, the AI Discovery Stack, the AI Provider Selection Pipeline, the Entity Corroboration Model, and the other published frameworks.
The historical record carries the pattern clearly.
FAQPage launched in May 2019 as a way to mark up frequently asked questions on dedicated FAQ pages. Within two years it had been adopted across millions of pages, most of which were not FAQ pages in any meaningful sense — they were product pages, service pages, and landing pages with FAQ blocks appended. Google narrowed FAQ rich result eligibility to government and health sites in August 2023. The remaining eligibility window was formally closed on 7 May 2026. Seven-year lifespan, decay driven by misuse at scale.
HowTo launched alongside FAQPage in 2019. It produced visually dominant SERP features — step-by-step rich results that displaced organic listings on procedural queries. Google deprecated HowTo rich results in September 2023 after similar industrial-scale misuse. Four-year lifespan, identical pattern.
Speakable launched in 2018 as a way to mark up content for voice assistant delivery. It never received meaningful consumer support outside Google News partners. It remains technically valid but commercially inert.
Article, Person, Organization, Product, Service, Event, Place — the foundational types that describe real-world entities — have remained continuously supported since their introduction more than a decade ago. They have been refined, extended, and integrated into Knowledge Graph infrastructure, but they have not been deprecated and there is no public signal that any of them are at risk.
The pattern in the failures is consistent. Each deprecated type derived its commercial value primarily from a SERP feature reward — a visual enhancement, a rich card, a structured snippet that earned click-through advantage. When that reward existed, adoption accelerated. When adoption reached the point where misuse outpaced legitimate use, the reward was withdrawn. The markup itself was never the problem; it was always the reward’s function as an incentive that produced the spam pattern.
The pattern in the survivors is also consistent. Article, Person, Organization, Product, Service, Event and Place describe entities that exist independently of any specific platform feature. Their commercial value to consuming systems is structural rather than presentational. An AI model trying to answer “who is Sean Mullins?” needs the Person schema to associate the name with the entity. A retrieval pipeline trying to ground a claim about a product needs the Product schema to disambiguate which product. These types are infrastructure. The systems that consume them have no incentive to deprecate them because their value is not extractive — it is referential.
From this pattern, a predictive principle: any schema type whose commercial value comes primarily from triggering a presentation feature is at structural risk over a multi-year horizon. Any schema type whose value comes from describing a stable real-world entity is durable.
The principle has practical force. It explains why the SEO industry should have anticipated FAQPage’s deprecation rather than treating it as a surprise. It predicts that any new structured data feature pitched as “critical for [acronym] visibility” with no underlying entity foundation has a half-life. It explains why the current GEO advice cycle promising 3.2x AI Overview presence for FAQ schema implementation is structurally identical to the 2021 advice cycle promising rich result advantage — same incentive, same predictable trajectory, same eventual outcome. And it suggests that the structured data investments that compound — that survive every platform cycle and become more valuable over time — are the ones anchored to entities, relationships, and verifiable claims rather than features.
The Half-Life Pattern is not the only useful lens for thinking about schema longevity. But it is a simple, falsifiable, predictive framework, and the conclusions it produces are commercially actionable: invest in entity-descriptive schema, treat feature-dependent schema as time-limited optionality, and design schema architecture for survivability rather than next-quarter visibility.
The FAQSection proposal — the subject of Part Three — is interesting partly because it appears to migrate FAQ markup from the fragile class back into the durable class. The current FAQ schema vocabulary failed the half-life test because it depended on a SERP feature for adoption. A clean replacement that describes a structural element of a page accurately, with no rich result reward attached, would not be subject to the same incentive dynamics. Whether the proposal succeeds in becoming standard is a separate question, addressed in Part Four. But its design is on the durable side of the line.
Part Three — The FAQSection Case Study: What Vocabulary Failure Looks Like
On 10 May 2026, Joost de Valk filed issue #4816 at the schema.org GitHub repository. The title: “Introduce FAQSection as a WebPageElement subtype for FAQ content that’s part of a page, not the whole page.” A companion issue, #4817, proposes a clean answer property on Question to replace acceptedAnswer for publisher-authored FAQs. Andrea Volpini, founder of WordLift, commented in support within forty-eight hours. The proposals are now in the standard community review process.
The technical content of the proposal is straightforward and is laid out in detail in the issue. The intellectual content — the diagnostic reasoning that produced the proposal — is more important for understanding why this matters beyond FAQ markup.
De Valk’s diagnosis has two parts.
The first is the mainEntity misrepresentation point. FAQPage was designed for pages whose primary purpose is question-and-answer content — help centres, regulator FAQ pages, government information sites. The schema’s mainEntity property is meant to point at the questions, because on a true FAQPage the questions are what the page is about. The much more common case in practice is a page about a product, an article, a service, or a piece of legislation, with an FAQ section attached. The questions there are part of the page, not the point of the page. Publishers who wanted to mark up that section had two choices: declare FAQPage with mainEntity pointing at the questions (which misrepresents what the page is), or skip the markup entirely. Most chose misrepresentation, because the rich result reward was sitting on the FAQPage side of the choice. The abuse pattern was partly structural — the vocabulary did not let publishers describe what their pages actually were, so they picked the closest available misrepresentation, and that misrepresentation scaled to industrial volumes.
The second is the acceptedAnswer origin point. The acceptedAnswer property that FAQPage uses was borrowed from a different problem domain — community Q&A sites like Stack Overflow, where users vote answers up and one rises to the top of the page. For publisher-authored FAQs nobody is accepting anything; the answer is simply the answer. The semantic mismatch was present at launch in 2019. Most working SEOs never noticed it because the property name was treated as a checkbox to complete rather than a semantic claim to evaluate.
Both diagnoses are credited to de Valk because both are genuinely his contribution to the FAQ schema conversation. They generalise beyond FAQ markup, which is what makes them worth surfacing here.
The wider pattern, of which FAQPage is one instance, is what might be called borrowed-property semantic drift. Schema.org has accumulated a vocabulary over fifteen years through incremental addition, and properties introduced for one problem domain have routinely been reused in adjacent domains where their original semantics no longer fit cleanly. acceptedAnswer is one example. author is another: originally modelling book and article authorship, it now annotates AI-generated tweets and machine-written summaries where the concept of authorship has frayed considerably. mainEntity itself is a third: introduced to disambiguate the primary topic of a page, it has been misused to game rich result eligibility on pages whose main entity is plainly something else. The list could be extended in several directions, and each direction is its own analysis — review property drift, knowsAbout drift, sameAs misuse on non-equivalent entities, datePublished on pages with no meaningful publication concept.
The pattern matters because it suggests the FAQ schema deprecation is not an isolated cleanup. It is the first visible instance of a vocabulary correction that the schema.org community will eventually need to apply to several other types. Borrowed-property drift accumulates quietly and breaks loudly. FAQPage broke at scale this month because the misuse was visible to Google and the cost of supporting the misuse exceeded the value of the rich result. Other drift patterns are not yet at that breaking point, but the underlying dynamic is identical.
The technical shape of the FAQSection proposal is exactly right for the durability question raised in Part Two. FAQSection is proposed as a subtype of WebPageElement, sitting alongside the existing WPHeader, WPFooter, WPSideBar, SiteNavigationElement, Table and WPAdBlock. None of those siblings have rich result rewards. They exist to describe structural elements of a page accurately. They are referenced via hasPart from the parent CreativeWork, which means a product page can declare itself as a Product with an FAQSection part, rather than misrepresenting itself as primarily a FAQPage. The Question children inside the FAQSection are referenced via hasPart in turn. The answer on each Question uses a clean answer property rather than the community-voting acceptedAnswer. The result is a graph that describes the page accurately: this is a WebPage, its main entity is a Product, and one of its parts is an FAQ section containing these questions and these answers.
Every claim in such a graph is literally true of the page. The page is a WebPage. Its main entity is the Product. The FAQ is a section within the page, not the page itself. That accurate description is the property the spec does not currently let publishers express — and the property that AI retrieval systems will increasingly need to disambiguate hybrid pages correctly.
{
"@context": "https://schema.org",
"@type": "WebPage",
"mainEntity": {
"@type": "Product",
"name": "Coffee Grinder X",
"offers": {
"@type": "Offer",
"price": "199.00",
"priceCurrency": "USD"
}
},
"hasPart": {
"@type": "FAQSection",
"name": "Frequently Asked Questions",
"hasPart": [
{
"@type": "Question",
"name": "What grind size is best for espresso?",
"answer": {
"@type": "Answer",
"text": "A fine grind, similar to table salt."
}
}
]
}
}The example follows the structure from issue #4816 with a product-page hybrid because the product-page case is the most common one in practice and the one where current FAQPage misuse is most visible. The same shape applies to LegalService pages with FAQ sections, Article pages with FAQ sections, and any other primary entity type with a hybrid FAQ block. That portability is part of what makes the proposal durable: it does not solve only the product-page case, it solves the entire category of pages whose primary purpose is something other than the FAQ that happens to be on them.
Part Four — How Schema.org Actually Works
The schema.org vocabulary is one of the few remaining open governance layers of the machine-readable web. Almost no mainstream SEO content explains how it functions, which is unfortunate because the answer is genuinely interesting and increasingly commercially relevant.
Schema.org was launched in June 2011 as a joint initiative of Google, Bing, Yahoo and Yandex. Its purpose is to provide a shared vocabulary for structured data on the web so that publishers can mark up content once and have it understood across multiple consuming systems. Governance is handled by a Steering Group with members from the founding companies plus W3C representation, and the day-to-day work happens through the public GitHub repository at github.com/schemaorg/schemaorg.
Anyone can file an issue. The proposal lifecycle, in broad terms, looks like this. An issue is opened. Community members — including representatives from large implementers like Yoast, WordLift, Schema App and the major search engines — comment on the proposal. If the discussion converges on a workable design, the Steering Group reviews the proposal, and if accepted it gets staged for a future schema.org release. Releases happen roughly monthly. A new type or property goes from the development site to the production site once it has passed review, and from that point onwards it is part of the standard vocabulary.
The historical precedent worth understanding is FAQPage itself. Issue #1723 was opened on 23 August 2017, titled “Introduce FAQPage as subtype (sibling?) of QAPage, for ‘Frequently asked questions’.” The proposal was filed by R.V. Guha, one of the original schema.org founders and the Steering Group chair, with backing from Google colleagues. It was accepted into the vocabulary, and Google began supporting FAQPage rich results in May 2019 — roughly twenty-one months from filing to consumer support. That timeline is unusually fast and reflects the proposal’s founder-backed origin; community-filed proposals without internal Google support typically take longer, sometimes substantially longer.
The point of walking through the precedent is to clarify what FAQSection #4816 is likely to face. The proposal is well-designed and has at least one notable industry endorsement (Volpini). De Valk has standing as a major implementer through Yoast SEO’s installed base. The structural argument is straightforward and additive — it does not require deprecating FAQPage, only complementing it with a more accurate alternative for the partial-FAQ case. None of these factors guarantee acceptance, but they are favourable starting conditions.
What acceptance into the vocabulary would mean, and would not mean, is the part that working SEOs most often get wrong. Acceptance into the schema.org standard is one event. Consumer support is several separate events, each made independently. Google could choose to honour FAQSection in its parsing pipeline without producing any associated rich result — given the May 2026 FAQ deprecation, this is the more likely Google outcome than any new visual feature. Bing might choose to honour it more visibly because Bing has continued to treat FAQ schema as a useful signal beyond Google’s position. AI providers — OpenAI through ChatGPT search, Anthropic through Claude with web access, Perplexity, Google through Gemini grounding, Microsoft through Copilot grounding via Bing — each decide independently whether to weight a new schema type in their retrieval and synthesis pipelines. Enterprise RAG systems make the same decision at the deployment layer, often based on whatever the indexing library they use happens to support at the version they are running.
This decoupling is the part of the governance picture most worth internalising. Schema.org acceptance, search engine support, and AI system honoring are three separate things. A type can be standard but unrewarded (Speakable since 2018). A type can be standard and rewarded with a SERP feature that later gets withdrawn (FAQPage, HowTo). A type can be standard, never receive a Google rich result, and still be commercially valuable because Bing, Copilot, and certain AI retrieval pipelines find it useful. The framing “will Google support this proposal” treats the question as binary in a way the underlying system has not been binary for several years.
The practical implication for any business with a stake in being machine-readable is that participation in the governance layer has compounding value. Filing comments on proposals that affect your sector, watching the issue tracker for vocabulary changes that touch your structured data, and engaging in the schema.org community group are activities with low cost and high optionality. The vocabulary that AI systems will use to interpret your business in 2028 is being shaped in 2026, and the people shaping it are the ones who show up.
Part Five — What the AI-Extraction Era Wants from Schema
The schema.org vocabulary was designed in 2011 for a web whose consuming systems were search engine indexes. The structural assumption was that publishers would mark up content, search engines would parse the markup, and parsed structured data would feed visible SERP enhancements. The publisher’s reward was presentation: a rich card, a star rating, a knowledge panel, a featured snippet. The system’s reward was query satisfaction: users got better answers, search engines retained query share.
That assumption no longer describes the systems consuming structured data in 2026.
Retrieval-Augmented Generation pipelines — the architecture underlying ChatGPT search, Perplexity, Claude with web access, Microsoft Copilot grounding, and Google’s AI Mode — do not consume structured data primarily to produce presentation enhancements. They consume it to identify entities confidently, to disambiguate which entity a query is about, to ground generated claims in retrievable sources, and to score the credibility of those sources before incorporating them into a synthesised answer. The publisher’s reward is no longer a visible SERP feature; it is citation eligibility and entity recognition inside answer generation.
This shift changes what well-structured schema is actually good for. A hybrid product page that accurately declares itself as a Product with an FAQSection part, with each Question and Answer cleanly typed, with the Organization marked up with verifiable sameAs links to Wikidata and Companies House, with the founding Person attached via foundingMember and corroborated through LinkedIn and Crunchbase, is not optimising for any SERP feature. None of those entities trigger a Google rich result on a product page. What the markup does is produce a graph that an AI retrieval system can parse with high confidence: the page is about this product, the product is offered by this organisation, the organisation is led by this person, and each of those entities is independently verifiable.
That graph is what makes the page citable. Citation is the new presentation reward. The mechanics are different. The rich card era was retrieval-then-render: the search engine retrieved a structured snippet and rendered it. The AI extraction era is retrieval-then-evaluate-then-generate: the system retrieves candidate sources, evaluates their credibility and entity confidence, and generates a synthesised answer that may or may not name the source. The schema markup’s job in the second model is to maximise the probability that the page is in the candidate set at all, that the entities on the page are recognised and disambiguated correctly, and that the system can confidently attribute claims to the page in the generated response.
This is consistent with what practitioner research is observing about AI citation behaviour. The University of Toronto study (September 2025, thirteen industries, consumer electronics dominant) found that 92.1% of Google AI Overview citations came from earned media rather than owned content. Muck Rack’s Generative Pulse analysis (July–December 2025, over a million links analysed) reported 82% earned media citation share across the same period. Seer Interactive’s 2025 conversion data put the click-through rate from AI Overview citations at 14.2% versus 2.8% for non-cited organic positions on the same page. The signals point in the same direction: AI systems are increasingly selecting based on cross-source signals rather than self-declared authority, and the value of being cited is materially higher than the value of being ranked in many query categories.
The strategic translation for schema architecture is that structured data has moved from being a presentation lever to being an extraction substrate. The markup that compounds in this regime is the markup that helps the system understand what the page is, who the page is by, what entities the page references, and what claims the page makes that can be verified against external sources. The markup that decays is the markup whose only purpose was to trigger a now-absent rich result.
This is the framework the existing CITATE standard already encodes at the page level — declarative opening, defined terms inline, statistics with named sources, attributable claims with date and author. The AI Discovery Stack five-layer model already maps how schema fits into the broader retrieval architecture. The AI Recommendation Pipeline already describes how the move from retrieval to recommendation depends on corroboration signals that schema alone cannot provide. The piece that has been missing from the public schema discussion is the underlying explanatory model: that AI systems are increasingly evaluating semantic credibility rather than syntactic validity, and that this evaluation is the mechanism through which both schema misuse and self-promotional content lose their commercial effectiveness over time.
Part Six examines that mechanism directly.
Part Six — Validity vs Credibility: How AI Systems Evaluate What to Trust
Two kinds of evaluation are increasingly diverging in how structured data is judged by consuming systems.
Syntactic validity is what schema validators check. The JSON-LD parses. The required properties are present. The types resolve to schema.org definitions. The Rich Results Test passes. Google’s Structured Data Linter, Schema App’s validator, Yoast’s built-in test — these are all syntactic validity checks. They evaluate the form of the markup, not the truth of the claims it makes. A page can validate at 100% and still be saying things about itself that are not true in any meaningful sense.
Semantic credibility is what AI retrieval systems increasingly evaluate. The markup’s claims are cross-referenced against entity databases, against external mentions of the same entity, against the page’s own content, against the broader corpus the system has indexed. The question being answered is not “is this markup formally correct” but “is this markup’s description of the world consistent with everything else the system knows.” A page that declares itself a FAQPage when it is plainly a product page can pass every syntactic check available. It fails the credibility check the moment any retrieval system compares the markup to the rendered content or to how the page is referenced externally.
The contrast between the two evaluation modes is sharper than most SEO content acknowledges:
| Traditional schema tooling checks | AI retrieval systems increasingly evaluate |
|---|---|
| Syntax validity | Semantic consistency |
| Property formatting | Cross-source corroboration |
| JSON-LD structure | Entity trust |
| Required fields | Contextual accuracy |
| Parser compatibility | Machine confidence |
The left column is what most SEOs still unconsciously optimise for. The right column is what AI systems are observably moving toward.
The Lily Ray Observation
On 12 May 2026, Lily Ray (VP, SEO & AI Search at Amsive; founder of Algorythmic) published an observation on LinkedIn that documents this shift in operation.
The query “best procurement software for small business” produced a Google AI Overview that named Precoro, Procurify, Tradogram, ControlHub and Airbase as the recommended options. Ramp’s page “Best Procurement Software for Small Business in 2026” was cited as a source contributing to the answer, but Ramp itself did not appear in the recommendation list. The query “best membership management software” produced an AI Overview naming Wild Apricot, Member365, MemberClicks, JoinIt, Fonteva, Circle and Mighty Networks. Outseta’s page “15 Best Membership Management Software Platforms [2026]” was cited; Outseta was absent from the recommendations. The query “best scheduling app for small business” produced an AI Overview naming Homebase, Calendly, Square Appointments, Deputy and Setmore. Rippling’s page “8 Best Scheduling Software for Small Businesses in 2025” was cited; Rippling was absent.
Three independent verticals, same behaviour. The self-promotional listicle was treated as a source of market intelligence — the model used it to identify competitors, learn the category, and understand the comparative landscape. But the publisher’s own self-recommendation was not honoured. Ramp’s claim to be best procurement software, Outseta’s implicit positioning as one of the top fifteen membership platforms, Rippling’s placement among the best scheduling software — all of them appear to have been discounted at the recommendation synthesis step.
Ray’s framing of the observation is the most parsimonious available: Google AI Overviews may now be applying a trust discount to self-recommendation, while continuing to use the underlying content for entity discovery and market mapping. Three datapoints across three verticals is not a controlled experiment and the explanation must remain probabilistic, but the pattern is consistent enough across independent queries to warrant strategic attention.
The strategic significance — and this is the synthesis layer worth pausing on — is that the same mechanism that broke FAQPage at scale appears to be operating in AI recommendation synthesis. In both cases, the system is increasingly separating what content declares about itself from what content can be trusted to assert. FAQPage declarations were taken at face value for years and then stopped being honoured when the misuse became visible. Self-recommendation claims appear to be undergoing the same transition. The schema-vocabulary story and the AI-recommendation story are two surface manifestations of one underlying shift: AI systems are increasingly separating declaration from trust allocation.
Four plausible explanations for the behaviour
Reasonable people can disagree on the precise mechanism producing Ray’s observation, and the honest treatment is to lay out the candidate explanations probabilistically rather than asserting one. Four are worth considering.
Conflict-of-interest dampening. The most likely explanation. Google may now recognise first-party self-ranking patterns and deliberately suppress them in recommendation synthesis. This makes particular sense in “best X” queries where AI-generated answers carry implicit endorsement risk. If Google were to repeat a vendor’s own claim that the vendor is best, the endorsement liability would sit with Google. Discounting the self-recommendation while retaining the entity discovery utility is the obvious risk-management posture.
Source diversification weighting. The system may simply be preferring independent corroboration, consensus across multiple sources, and multi-source recurrence over single-source assertion. Being mentioned across Reddit, review sites, independent blogs, analyst reports, and competitors’ comparison pages becomes structurally more important than publishing your own “we’re best” page. This aligns with the broader Entity Corroboration pattern documented in the AI Provider Selection Pipeline framework.
Entity extraction without recommendation trust. The system may use self-promotional listicles for entity discovery and category understanding while declining to trust the same source for recommendation authority. The page contributes to entity awareness, category mapping and competitive comparison without influencing the “who should I recommend” output. This is the most architecturally specific reading of what Ray’s screenshots appear to show.
Temporary anti-spam adjustment. Also possible. Google may simply be reducing obvious manipulation vectors during a tuning phase, with the weighting potentially evolving again later. Treating the observation as a stable feature of the system rather than a current configuration is premature; the honest reading is that something is being tested and the direction is consistent with the broader credibility-evaluation shift.
Whichever explanation turns out to be most accurate, the strategic implication is similar in all four cases.
What this implies for the AI-readable web
If AI systems are increasingly separating entity discovery from trust allocation — whether through deliberate conflict-of-interest suppression, source diversification weighting, or some combination — four directional consequences follow.
Independent corroboration grows in importance. Mentions across analyst commentary, third-party review sites, Reddit and other community platforms, directories, expert blogs, partner ecosystems, and media coverage become structurally more important for recommendation inclusion than any volume of self-published assertion. The Entity Corroboration Model already names this; the AI Overview observation is one of the cleanest empirical signals it has produced.
Consensus synthesis strengthens. AI systems are likely to keep moving toward “what do multiple independent sources consistently say” as the dominant evaluation question, rather than “what does this page claim.” This is a substantial conceptual shift from the ranked-list logic that dominated the search era. The unit of evaluation is becoming the entity’s reputation across the web rather than any individual page’s claims about it.
Self-promotional comparison content loses recommendation effectiveness but retains entity discovery utility. The strategic implication for content teams is that “best X” listicles published by vendors in category X are not useless — they appear to still earn citation and contribute to category mapping. But they no longer reliably translate into recommendation inclusion, which means the implicit ROI assumption underlying most vendor-published listicles is being eroded. The content’s job is shifting from “get us recommended” to “teach the model about our category and our competitors.” That is still useful work, but it is a different brief and a different success metric.
Semantic trust signals matter more. Authorship, citations, corroborated identity, machine-readable entity infrastructure, and behavioural consistency across sources all become more commercially valuable. CITATE’s six-criterion structure — clarity, intent architecture, trust signals, attribution, transparency, evidence — describes the page-level expression of this shift. The Entity Corroboration Model describes its off-page expression. Both are converging on the same underlying property: semantic credibility, not syntactic validity, is the standard structured data must now meet.
The five-question test for any structured data implementation in 2026 is no longer just “does it validate.” It is closer to this:
- Does the @type accurately describe what the page actually is, or is it the closest available misrepresentation?
- Do the properties match their original semantic intent, or have they been borrowed from a different problem domain and forced to fit?
- Are the entities the markup names corroborated externally — Wikidata, Companies House, LinkedIn, editorial coverage — or only declared on the page itself?
- Is the markup readable by an extraction system trying to ground a claim, not just by a crawler trying to render a snippet?
- Would the markup survive the Half-Life test — does it describe a stable real-world entity, or is its value entirely tied to a current platform reward?
A page that passes all five is doing structured data in the way the AI-extraction era rewards. A page that passes the validator but fails three or more of these is doing structured data in the way the SERP-feature era used to reward. The first kind of page compounds. The second kind has a half-life.
Part Seven — What This Means For Your Site Now
The strategic implications of the validity-to-credibility shift translate into a small number of practical actions.
Audit your existing schema against the Half-Life Pattern. Inventory every page on your site that emits structured data. Categorise each markup block by whether it describes a stable real-world entity (Person, Organization, Product, Service, Article, Event, Place — durable) or whether its value is primarily tied to a platform feature (FAQPage, HowTo, Speakable — fragile). For the fragile category, evaluate whether the markup is still honest given the current state of the page. FAQ markup on a true help centre or regulator FAQ page is fine. FAQ markup on a service page footer where the FAQ is decorative is the misrepresentation pattern that drove the May 2026 deprecation and should be removed regardless of whether Google still parses it. The cost of removing decorative misrepresentation is zero. The cost of keeping it is that AI extraction systems increasingly treat its presence as a credibility signal in the wrong direction.
Build entity infrastructure before you optimise pages. The corroboration signals that matter for AI recommendation eligibility are mostly off-page rather than on-page. Wikidata entry. Companies House record. Crunchbase profile. LinkedIn organisation page with consistent NAP. Apple Business Connect. Editorial coverage that names your business in the specific category you want to be recommended for. Each of these is a verification node that allows an AI retrieval system to cross-reference what your structured data claims against an independent source. The page-level schema is the on-page expression of an entity that needs to exist independently across multiple databases to be retrievable with confidence.
Distinguish content built for entity discovery from content built for recommendation inclusion. The Lily Ray observation suggests these are increasingly two different jobs. Self-published comparison pages, category overviews, and educational content about your sector are useful for entity discovery and category mapping — AI systems use them to understand who exists in your space and how the category is structured. They are not reliably useful for getting yourself recommended. Recommendation inclusion appears to depend on corroboration from sources you do not control: editorial coverage, independent reviews, community discussion, analyst commentary, and the long tail of mentions across platforms that AI systems consult when synthesising a recommendation. Both kinds of content are worth producing, but the success metrics for them are different and the misallocation of effort between them is one of the more common strategic errors in AI-era SEO.
Treat schema participation as governance, not just implementation. The schema.org community group is open. Issues are visible. Comments matter. If your sector has structured data needs that the current vocabulary does not handle cleanly, filing an issue or commenting on existing proposals is genuinely low-cost activity with compounding optionality. The vocabulary that AI systems will use to interpret your business in 2028 is being shaped in 2026, and the practitioners who participate at the governance layer are the practitioners whose concerns get represented in the eventual standard.
Apply the five-question test to new structured data implementations. Before deploying any new schema markup, run it through the Part Six checklist. Does the @type accurately describe what the page actually is. Do the properties match their original semantic intent. Are the entities corroborated externally. Is the markup readable by an extraction system not just a crawler. Would the markup survive the Half-Life test. Markup that passes all five compounds in the AI-extraction era. Markup that fails three or more is structurally fragile regardless of how cleanly it validates.
Forward reading from here:
- CITATE — the page-level standard for AI citation readiness, applied here as the structural expression of semantic credibility at the page level.
- The AI Discovery Stack — the five-layer model showing how schema fits into the broader retrieval architecture and where structured data sits relative to entity infrastructure and recommendation eligibility.
- Entity Corroboration — the off-page expression of the credibility shift, detailing how external verification signals translate into AI retrieval confidence.
- FAQ Schema Deprecation 2026 — the focused decision-tree-and-action-plan piece for the specific May 2026 FAQ rich result deprecation event, the originating context for this guide.
- AI Visibility Audit — the consultancy engagement that applies these frameworks to a specific business’s structured data, entity infrastructure, and corroboration posture.
Closing
The web is moving from machine-readable to machine-verifiable. The two recent developments that opened this guide — FAQ rich result deprecation and the AI Overview self-recommendation discount — are early visible signals of a shift that is going to define the structured data conversation for the rest of this decade. Schema validity is increasingly insufficient on its own. Semantic credibility is the standard structured data must now meet to retain its commercial value, because the systems consuming structured data are increasingly evaluating credibility directly rather than treating syntactic correctness as a proxy for it.
For working SEOs and content teams, the implication is that schema architecture in 2026 is no longer a technical specialism that lives in a corner of the discipline. It is the page-level expression of how your business will be understood by AI systems for the next decade, and the schema decisions made now will compound or decay according to whether they describe reality accurately or just trigger features.
The vocabulary is being remade in public. Whether your industry’s schema reflects reality accurately enough for AI systems to trust it is now a strategic question, not a niche curiosity.
